Mastro

Version unknown · Measured 6/29/2026

Dev Time Performance

Prod Deps	Dev Deps	Dup. Deps	node_modules	node_modules (prod)	Dep Install Size	Graph
2	2	0	29.16MB	3.36MB	27.39MB	View

Metric	Avg	Min	Max
Install	0.53s	0.49s	0.56s
Cold Build	0.37s	0.36s	0.38s
Warm Build	0.36s	0.36s	0.37s

Build output size: 0.01MB

Framework	First Paint	FCP	INP

Each framework renders a table of 1000 rows with two UUID columns
Measured using Lighthouse flow with Chromium via Puppeteer for accurate browser metrics
First Paint and First Contentful Paint are measured on initial navigation
Interaction to Next Paint is measured by clicking the first row's detail link
Benchmarks run 5 times and results are averaged
Next.js wraps the client-side rendered table in a dynamic import with ssr: false to prevent build-time prerendering
TanStack Start, Nuxt, SvelteKit, and SolidStart disable SSR per-route
React Router uses route-level clientLoader functions with HydrateFallback so the client-rendered routes are not server-rendered
Astro uses client-only React islands for client-side rendered routes
Client-side rendered tests use each framework's normal production build because SPA-only build modes are not supported consistently across the frameworks being compared
Astro uses React for its client-side rendered test: the benchmark table and detail components are React islands rendered with client:only="react", which prevents Astro from server-rendering those components and lets them render only in the browser. Astro's ClientRouter is not used for this CSR test because it enables client-side transitions and soft navigation behavior rather than client-only rendering.

Framework	First Paint	FCP	INP

Each framework renders a table of 1000 rows with two UUID columns
Measured using Lighthouse flow with Chromium via Puppeteer for accurate browser metrics
First Paint and First Contentful Paint are measured on initial navigation
Interaction to Next Paint is measured by clicking the first row's detail link
Benchmarks run 5 times and results are averaged
The measured route is /server-side-rendered, and detail navigation uses /server-side-rendered/:id.

Framework	Ops/sec	Median Latency	Body Size	Duplication
Baseline HTML	841	1.206ms	96.83kb	1x
Mastro	515	1.912ms	181.95kb	1x

Each framework renders the dedicated /ssr-throughput route with a table of 1000 rows and UUID id/name columns
This route intentionally does not render the exact same table as the browser SSR and load tests: it omits detail links and framework link components so router, prefetch, and navigation metadata do not dominate the request-handler throughput measurement
Mock HTTP requests bypass TCP overhead so this measures request-handler rendering throughput rather than full server throughput
Data is loaded asynchronously to simulate real-world data fetching
Duplication factor indicates how many times each UUID appears in the response (1x = optimal, 2x = includes hydration payload)
Benchmarks run for 10 seconds using tinybench
Frameworks are invoked through their production request handlers where possible. Web API handlers are called with Request objects; Node.js handlers are called with mock IncomingMessage and ServerResponse objects.
Next.js renders the throughput table as a client component, matching the setup from PR #94, so the benchmark compares traditional server-rendered React + hydration work instead of making Next.js render every table row as React Server Components
Inspired by eknkc/ssr-benchmark

Framework	Peak req/s	Peak Connections	P99 @ 25	P99 @ 50	P99 @ 100	Total Req.
Baseline HTML	1,629.8	50	20ms	44ms	107ms	48,210

Each framework serves the server-rendered table route over a real local HTTP server
The measured route is /server-side-rendered, using the same 1000-row UUID table as the SSR request throughput and browser rendering tests
Load is applied in staged connection counts, from 1 through 200 concurrent connections, with each stage running for approximately 5 seconds
Peak requests/sec is the highest successful stage throughput observed during the staged run
P90 and P99 latency are compared at the 25-, 50-, and 100-connection stages for every framework, so latency is measured under the same concurrency pressure
Total requests cover the full staged load run, not only the peak stage